Recognizing an apple is a simple thing for even young toddlers. It’s not for a drone, which needs to train its AI models on many, many pictures of apples: An apple in the shade, an apple in the sun, an apple from the front, from the back, from a frog’s perspective…
It takes time, but machine learning effectively learns the same way humans do. “For a baby to know that this is an apple, it can look at different kinds of pictures and recognize that this is an apple”, says Andy Kong, a research assistant at the University of Hong Kong.
Microsoft and DJI say that drones powered with this technology can be used in a number of scenarios. If you own huge warehouse, drones can help reach areas inaccessible to workers and check stock. If you’re a real estate company you could easily inspect buildings for cracks and rust.
That’s why the two companies organized the AI x Drones Joint University Competition, gathering students from three Hong Kong universities to make AI models which help drones recognize different types of fruit.
The students then released their drones in areas with fruit scattered around, and left them to do their magic. Screens displayed how successful their models were -- labeling the fruit, plus a number indicating how sure it is that the banana is indeed a banana.
The teams used a DJI Phantom 4 Pro, capable of shooting 4K ultra HD footage -- which Microsoft’s AI then scrutinizes to figure out what it’s seeing.
The secret weapon here? Analysing data close to its source. Cloud computing is more powerful, sure, but the transmission time adds a huge lag.
